[WIP] Add PyTorch-Warp interoperability in Chapter 12 notebook #75
Conversation
Accelerated_Python_User_Guide/notebooks/Chapter_12_Intro_to_NVIDIA_Warp.ipynb
Outdated
Show resolved
Hide resolved
Accelerated_Python_User_Guide/notebooks/Chapter_12_Intro_to_NVIDIA_Warp.ipynb
Show resolved
Hide resolved
Accelerated_Python_User_Guide/notebooks/Chapter_12_Intro_to_NVIDIA_Warp.ipynb
Outdated
Show resolved
Hide resolved
Accelerated_Python_User_Guide/notebooks/Chapter_12_Intro_to_NVIDIA_Warp.ipynb
Outdated
Show resolved
Hide resolved
Accelerated_Python_User_Guide/notebooks/Chapter_12_Intro_to_NVIDIA_Warp.ipynb
Outdated
Show resolved
Hide resolved
Accelerated_Python_User_Guide/notebooks/Chapter_12_Intro_to_NVIDIA_Warp.ipynb
Outdated
Show resolved
Hide resolved
Accelerated_Python_User_Guide/notebooks/Chapter_12_Intro_to_NVIDIA_Warp.ipynb
Outdated
Show resolved
Hide resolved
Accelerated_Python_User_Guide/notebooks/Chapter_12_Intro_to_NVIDIA_Warp.ipynb
Outdated
Show resolved
Hide resolved
| "cell_type": "markdown", | ||
| "metadata": {}, | ||
| "source": [ | ||
| "Let us define the backward function for our optimization and register it using `register_autograd` for our `warp_rosenbrock` operator" |
There was a problem hiding this comment.
I guess there's sort of a plummet in quality after this and we just have the reader execute code block after code block, not expecting if they should see anything after executing one or what the code block is doing. Are you able to improve the flow (from your perspective as someone who probably hasn't seen this example before)?
Accelerated_Python_User_Guide/notebooks/Chapter_12_Intro_to_NVIDIA_Warp.ipynb
Show resolved
Hide resolved
| "nbconvert_exporter": "python", | ||
| "pygments_lexer": "ipython3", | ||
| "version": "3.12.3" | ||
| "version": "undefined.undefined.undefined" |
There was a problem hiding this comment.
Not sure how it got to be "version": "undefined.undefined.undefined"
There was a problem hiding this comment.
I am not sure what happened there. In my current push, it shows 3.11.11
| "metadata": {}, | ||
| "outputs": [], | ||
| "source": [ | ||
| "# Our backward functions is warp_rosenbrock_backward, being wrapped def backward(...) wraps it\n", |
There was a problem hiding this comment.
- functions -> function
- warp_rosenbrock_backward ->
warp_rosenbrock_backward - "being wrapped def backward(...) wraps it" is kind of a nonsense comment
There was a problem hiding this comment.
Lol, what was I thinking writing that comment in the third bullet, fixed now
| ] | ||
| }, | ||
| { | ||
| "cell_type": "code", |
There was a problem hiding this comment.
Still need a markdown cell before each block of code that explains what the following code block does
Accelerated_Python_User_Guide/notebooks/Chapter_12_Intro_to_NVIDIA_Warp.ipynb
Outdated
Show resolved
Hide resolved
Accelerated_Python_User_Guide/notebooks/Chapter_12_Intro_to_NVIDIA_Warp.ipynb
Outdated
Show resolved
Hide resolved
| "cell_type": "markdown", | ||
| "metadata": {}, | ||
| "source": [ | ||
| "In the GIF above, the red star represents the optimum, the white dot shows the mean of $(x, y)$ coordinates across all 1500 points, and the black dots are the individual points. As the optimization progresses, both the white dot and the individual black dots converge toward the red star, validating our hybrid PyTorch-Warp optimization scheme for the Rosenbrock function." |
There was a problem hiding this comment.
There should be some acknowledgement here about the simplicity of the example. The PyTorch custom operator was extremely simple and this optimization problem can be easily done in PyTorch without going to a kernel-based framework like Warp. We should have the reader recognize that instead of simplify evaluating a function, their custom operators might contain entire simulators with multiple kernel launches that might also use Warp's built-in spatial data structures.
3 participants
Added a section on PyTorch-Warp interoperability towards the end of Chapter 12 notebook.